IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 10, NO.

2, FEBRUARY 2011 495

Blind Channel Estimation for STBC Systems
Using Higher-Order Statistics
Vincent Choqueuse, Member, IEEE, Ali Mansour, Senior Member, IEEE, Gilles Burel, Senior Member, IEEE,
Ludovic Collin, and Kof Yao, Member, IEEE
AbstractThis paper describes a new blind channel estimation
algorithm for Space-Time Block Coded (STBC) systems. The
proposed method exploits the statistical independence of sources
before space-time encoding. The channel matrix is estimated
by minimizing a kurtosis-based cost function after Zero-Forcing
equalization. In contrast to subspace or Second-Order Statistics
(SOS) approaches, the proposed method is more general since it
can be employed for the general class of linear STBCs including
Spatial Multiplexing, Orthogonal, quasi-Orthogonal and Non-
Orthogonal STBCs. Furthermore, unlike other approaches, the
method does not require any modication of the transmitter
and, consequently, is well-suited for non-cooperative context. Nu-
merical examples corroborate the performance of the proposed
algorithm.
Index TermsMIMO, space time coding, channel estimation,
independent component analysis, higher-order statistics.
I. INTRODUCTION
S
PACE-TIME Block Coding is a set of practical signal
design techniques aimed at approaching the information
theoretic capacity limit of Multiple-Input Multiple-Output
(MIMO) channels. Since the pioneer work of Alamouti [1],
space-time coding has been a fast growing eld of research.
In the last decade, numerous coding schemes have been
proposed. These include orthogonal (OSTBCs) [1][3], quasi-
orthogonal (QOSTBCs) [4], [5] and non-orthogonal STBCs
(NOSTBCs) [6]. At the receiver side, the decoding is achieved
by a space-time equalizer. Most space-time equalizers require
Channel State Information (CSI). This information is usually
obtained through training based techniques at the expense of
the bandwidth efciency. On the other hand, the differential
schemes proposed in [7][10], which do not require CSI, incur
a penalty in performance of at least 3dB as compared to the
coherent Maximum-Likelihood (ML) receiver. The drawbacks
of training-based approaches and differential schemes have
motivated an increasing interest in the development of blind
Manuscript received October 23, 2009; revised March 2, 2010 and October
1, 2010; accepted October 18, 2010. The associate editor coordinating the
review of this paper and approving it for publication was L. Deneire.
V. Choqueuse was with the Lab-STICC, UEB, Universit de Brest; UMR
CNRS 3192, CS 93837, 29238 Brest cedex 3, France. He is now with LBMS,
EA 4325, same address (e-mail: vincent.choqueuse@gmail.com).
A. Mansour is with the Department of Electrical and Computer Engineer-
ing, Curtin University, Perth, Australia (e-mail: ali.mansour@ieee.org).
G. Burel, L. Collin, and K. Yao are with the Lab-STICC, UEB, Universit
de Brest; UMR CNRS 3192, CS 93837, 29238 Brest cedex 3, France (e-mail:
{Gilles.Burel, Ludovic.Collin, Kof-Clement.Yao}@univ-brest.fr).
The authors would like to thank Janet Leschaeve for her help in checking
English usage, as well as the associate editor and the anonymous reviewers
for their useful comments.
Digital Object Identier 10.1109/TWC.2010.112310.091576
channel estimation algorithms for STBC systems. Develop-
ment of blind receivers also has applications in military
communication system when the transmitted symbols have to
be estimated in a blind fashion.
Blind channel estimation algorithms based on Maximum-
Likelihood (ML) have been proposed in [11], [12]. Despite
their high performances, the computational costs of the ML-
based methods become prohibitive for high-order modulations.
In the case of BPSK or QPSK constellations, the blind-ML
detection can be simplied to a Boolean Quadratic Program
(BQP) [13]. For more general settings, iterative procedure
can be employed to avoid the computational complexity of
the ML approach. These include the Cyclic ML [12] and the
Expectation-Maximisation (EM) [14], [15] algorithms. How-
ever, these iterative methods require a careful initialization of
the channel and/or symbols. In particular, a poor initialization
can strongly affect the Symbol-Error Rate (SER) performance.
To avoid the drawbacks of ML-based channel estimation
algorithms, several authors have investigated the use of sub-
space [16], [17] or Second-Order Statistics (SOS) [18][21]
approaches. However, excluding some specic low-rate codes,
these approaches fail to extract the channel in a full-blind
context [16][22]. Several approach have been proposed in
literature to solve this problem, including the transmission of a
short training sequence [16], [17] or the use of precoders [18],
[20], [21]. However, these semi-blind methods cannot be
employed in a non-cooperative scenario since they require
modication of the transmitter.
One solution to avoid the limitations of SOS and subspace
algorithms is to exploit Higher-Order Statistics (HOS). This
approach is usually called Independent Component Analysis
(ICA) [23]. ICA was originally developed for non-coded
systems. Recently several authors have investigated its ex-
tension to STBC communications [24][31]. Nevertheless,
these extensions have several limitations and drawbacks. In
particular, the algorithms [24][29] are limited to a sub-class
of Orthogonal STBCs and their extension to the general class
of STBCs is far from trivial. On the other hand, the methods
[30], [31] do not take into account the specic structure of the
STBC.
Despite this rich literature, none of the previous algorithms
is able to estimate the channel matrix for general STBCs with-
out modication of the transmitter (pilot sequence, precoding).
In this paper, an original algorithm is proposed which is well-
suited to the general class of linear STBCs whatever the code-
rate and/or the modulation. The channel matrix is estimated
1536-1276/11$25.00 c 2011 IEEE
h
t
t
p
:
/
/
w
w
w
.
i
e
e
e
x
p
l
o
r
e
p
r
o
j
e
c
t
s
.
b
l
o
g
s
p
o
t
.
c
o
m
h
t
t
p
:
/
/
w
w
w
.
i
e
e
e
x
p
l
o
r
e
p
r
o
j
e
c
t
s
.
b
l
o
g
s
p
o
t
.
c
o
m
496 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 10, NO. 2, FEBRUARY 2011
by minimizing a kurtosis-based cost function. In contrast with
classical ICA algorithms, the cost-function is computed from
the Zero-Forcing (ZF) space-time equalized symbols. Our
proposed method has low-complexity and does not require
any channel initialization, the use of pilot sequence and/or
knowledge of modulation. Furthermore, our method does not
require any modication of the transmitter and, consequently,
can be employed in a non-cooperative scenario.
This paper is organized as follows. The signal models and
the assumptions are presented in section II. The kurtosis-based
cost function is described in section III and the minimization
algorithm is described in section IV. The set of the remaining
ambiguities after channel estimation is provided in section V.
Finally, the simulation results are presented in section VI.
II. SIGNAL MODELS AND ASSUMPTIONS
Hereinafter, bold upper case letters denote matrices, e.g.,
X; bold lower case letters stand for column vectors, e.g., x,
and lower case letters represent scalars. Superscripts ()
T
and
()
H
denote transpose and Hermitian, respectively. Symbol
=

1 is the imaginary unit, ()

corresponds to the
complex conjugate and the operators () and () denote
the real and imaginary parts, respectively. The symbol is
the Kronecker product, trace() is the trace function and []
is the expectation operator. The matrices I

and 0

correspond to the identity and zero matrices, respectively. The

unit vector, e
()

, is an -dimensional row vector with "1" in

its

component and zero elsewhere i.e.

e
()

= 0 0 1 0 0
(1)
The elementary matrix, E
()

, is an matrix which is "1"

in the

row and

column and zero elsewhere i.e.

E
()

.
.
.
0
0 1 0 . . .

0
.
.
.
(2)
A. Transmitted signal model
Let us consider a linear STBC that transmits symbols
during time slots through

antennae. The space-time block

encoder generates an

block matrix from a block of

symbols s = [
1
, ,

]
T
. The block matrix, (s), can be
expressed under the general form [32]
(s) =

=1

) + A
+
(

(3)
where the

matrices A

are the space-time coding

matrices.
B. Signal model of received samples
Let us consider a receiver composed of

antennae. Let
us also assume a quasi-static frequency-at channel modelled
by an

complex matrix H. The

received block,
denoted by the

matrix Y

, is given by [32]
Y

= H(s

) + B

(4)
where the

matrix B

= [b

(1), , b

()] refers to the

additive noise and b

() is a

-dimensional column vector.

The aim of this study is to estimate H from the received data
blocks, Y

, under the following assumptions:

AS1) the

channel matrix, H, is of full-column

rank. Furthermore, the number of receiver antennae is
strictly greater than the number of transmitters, i.e.

>

.
AS2) the noise vector is both spatially and temporally
white with a variance of
2
per complex dimension. In
particular, it implies that:
[B

B
H

] =
2
I

(5)
AS3) the transmitted symbols, s

, are non gaussian,

independent and identically distributed (i.i.d).
AS4) the average transmit power on each antenna is
normalized to unity which also implies that:
[(s

)
H
(s

)] = I

. (6)
AS5) the space-time code is known at the receiver side.
Assumptions AS1), AS2) and AS3) are widely used and
AS4) is respected for most STBCs
1
. Moreover in many sce-
narios, the space-time code is usually assumed to be known,
otherwise, it can be estimated with a blind STBC recognition
algorithm [33][36]. It should be noted that condition AS5)
also implies that , ,

and A

are known at the receiver

side.
III. CHANNEL ESTIMATION STRATEGY
In this section, a new blind channel estimation strategy
based on HOS is proposed. The method is composed of two
steps which are detailed in the subsections III-A and III-B,
respectively.
A. Step 1: Data Whitening
In the preprocessing step, the channel is estimated up to a
unitary matrix through the use of SOS. By using assumptions
AS1), AS2), AS3) and AS4), the

covariance matrix
of the noiseless transmitted signals R = [Y

Y
H

]
2
I

can be expressed as
R = H[(s

)
H
(s

)]H
H
= HH
H
. (7)
Under assumption AS1), the rank of the symmetric matrix R
is equal to

. Therefore, R can be decomposed as follows:

1
It should be noted that if [(s

)
H
(s

)] = I

, the scaling factor

can be absorbed into the channel matrix H without loss of generality
h
t
t
p
:
/
/
w
w
w
.
i
e
e
e
x
p
l
o
r
e
p
r
o
j
e
c
t
s
.
b
l
o
g
s
p
o
t
.
c
o
m
h
t
t
p
:
/
/
w
w
w
.
i
e
e
e
x
p
l
o
r
e
p
r
o
j
e
c
t
s
.
b
l
o
g
s
p
o
t
.
c
o
m
CHOQUEUSE et al.: BLIND CHANNEL ESTIMATION FOR STBC SYSTEMS USING HIGHER-ORDER STATISTICS 497
R = UU
H
, where U is an

matrix satisfying U
H
U =
I

and is an

diagonal matrix containing real entries.

From (7), it follows that the channel matrix Hcan be expressed
as
H =
1

U
1
2
W
H
(8)
where W is an

full rank unitary matrix.

After the preprocessing step, the determination of the matrix
H reduces to the determination of the

unitary matrix
W. To determine W, let us dene the

whitened data
block, X

, as
X

1
2
U
H
Y

. (9)
B. Determining the unitary matrix W
1) The Zero-Forcing receiver: In this paragraph, a linear
Zero-Forcing (ZF) decoder is expressed in terms of the

whitened data block X

. Using (4) and (8), it can be shown

that
X

= W
H
(s

) +

1
2
U
H
B

. .. .
N

(10)
where the

matrix N

is a multidimensional zero-mean
Gaussian signal. Let us dene, s

, the 2 real-valued column

vector obtained by concatenating the real and imaginary part
of s

i.e.
s

(
()
1
)
.
.
.
(
()

)
(
()
1
)
.
.
.
(
()

. (11)
Let us also introduce the vectorization operator, vec{.}, ob-
tained by stacking all columns of a matrix on top of each
other. Using the property of the vec{.} operator in equations
(10) and (3) [37], it can be shown that
x

= WGs

+ n

(12)
where the 2

-dimensional column vectors x

and n

, the
2

matrix W and the 2

2 matrix G are
respectively dened by
x

(vec{X
H

})
(vec{X
H

})

(13)
n

(vec{N
H

})
(vec{N
H

})

(14)
G

(vec{A
H
1
}) (vec{A
H
2
})
(vec{A
H
1
}) (vec{A
H
2
})

(15)
W

(W
T
) I

(W
T
) I

(W
T
) I

(W
T
) I

. (16)
As W is a unitary matrix, it is demonstrated in appendix A
that W is orthogonal i.e. W
T
W = I
2

. If the unitary matrix

W is known at the receiver side, the transmitted symbols can
be recovered with a linear Zero-Forcing (ZF) equalizer. The
ZF equalizer computes an inverse matrix to compensate the
combined effects of the channel and space-time coding i.e.

(
()
1
)
.
.
.
(
()

)
(
()
1
)
.
.
.
(
()

= G

(W) I

(W) I

(W) I

(W) I

. .. .
W
T
x

(17)
where the 22

matrix G

denotes the pseudo-inverse of

G (G

G = I
2
) and
()

is the

estimated symbol of the

block. Using (17),

()

can be expressed as

()

= [e
()

e
()

]G

W
T
x

. (18)
In a blind context, the unitary matrix W is unknown at the
receiver side. To estimate W, this study exploits the statistical
independence of the equalized symbols. More precisely, the
unitary matrix W is estimated by maximizing the statistical
independence of the Zero-Forcing equalized symbols,
()

.
2) Kurtosis-based cost function: A simple approach to
maximize the statistical independence of
()

is to maximize
the nongaussianity of
()

[23]. One measure of nongaussian-

ity of a random variable is the (unnormalized) Kurtosis,
[], which is dened as
[] [
4
] 2([
2
])
2
[][

]. (19)
It follows that the unitary matrix W can be estimated by
maximizing the function

=1
[
()

] where . denotes the

absolute value. It should be noted that in the most practical
cases, the sign of the kurtosis is assumed to be known and
the same for all the transmitted symbols. In particular, it is
shown in reference [38] that the cumulant of most of the digital
modulation (ASK, PSK and QAM) are negative. Therefore, an
estimate of W, denoted

W, can be obtained as follows

W :

min
W
(W) =

=1

()

subject to WW
H
= I

(20)
where (W) is a real-valued cost function which depends
on the

complex-valued matrix W. It should be

noted that criterion (20) has already appeared in literature for
simpler channel estimation problems. In particular, it has been
employed for classical ICA problems, where (s

) = s

[39]
[42]. In our study, an extension to STBC systems is obtained
by applying criterion (20) on the Zero-Forcing space-time
equalized symbols
()

in (18).
IV. ALGORITHM IMPLEMENTATION
In this section, the focus is on the minimization of the real-
valued cost function :

under the unitary

constraint WW
H
= I

. As no closed form solution exists,

a Steepest-Descent (SD) approach is employed. To perform
a descent step, SD algorithm requires the computation of
the gradient. The gradient expression has been provided in
several studies for classical ICA problems [42], [43], however,
h
t
t
p
:
/
/
w
w
w
.
i
e
e
e
x
p
l
o
r
e
p
r
o
j
e
c
t
s
.
b
l
o
g
s
p
o
t
.
c
o
m
h
t
t
p
:
/
/
w
w
w
.
i
e
e
e
x
p
l
o
r
e
p
r
o
j
e
c
t
s
.
b
l
o
g
s
p
o
t
.
c
o
m
498 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 10, NO. 2, FEBRUARY 2011
its expression is no longer valid for STBC systems. In the
subsection IV.IV-A, the gradient expression is established
for STBC systems. Then, two constrained-minimization SD
algorithms are described in subsection IV.IV-B.
A. Expression of the gradient in the Euclidean space
In the Euclidean space, the gradient of the cost function
(W) is the

matrix
W
which is dened as [44]

W
=
(W)
W

(21)
where:
(W)
W

1
2

(W)
(W)
+
(W)
(W)

. (22)
Let us denote,

, the element on the

row and the

column of matrix W. Using the

elementary matrix
E
(

,
W
can be expressed as

W
=

=1

=1
1
2
E
(

(W)
(

)
+
(W)
(

. (23)
From (20), it follows that

W
=

=1

=1

=1
1
2
E
(

[
()

]
(

)
+
[
()

]
(

(24)
where

is given by (18). By interchanging the order of

derivative and expectation [45], the derivative of the (unnor-
malized) Kurtosis with respect to a complex element is given
by
[

= 2

[
2

+[
2

. (25)
From (18), the

matrix
W
can be expressed as

W
=

=1

=1

=1
E
(

()

()

q
(2)

+
()

q
(1)

()

()

q
(2)

+
()

q
(1)

[
2()

()

q
(2)

[
2()

()

q
(1)

(26)
where the 2

-dimensional column vector x

is dened in
(13) and where the 2

-dimensional row vectors q

(1)

and
q
(2)

are given respectively by

q
(1)

= [e
()

e
()

]G

E
(

E
(

E
(

E
(

(27)
q
(2)

= [e
()

e
()

]G

E
(

E
(

E
(

E
(

. (28)
Remark 1: In practice, the signals are assumed to be er-
godic; that means that the expectation operator [] in (26)
can be approximated by a time-average.
B. Constrained minimization algorithm
Several SD algorithms for the minimization of a real-valued
cost function under the unitary constraint have been proposed
in literature. In this subsection, two algorithms are described.
For constrained-minimization, classical approaches solve
the optimization problem on the Euclidean space by using
gradient-based algorithms [23], [42], [46]. At each iteration
step, an update of W is performed in the direction of the
negative gradient. Then, a symmetric orthogonalization
is applied to restore the unitary constraint of W. This
two-step approach is described in the algorithm 1 for a
xed step size
2
. The major drawback of the Euclidean SD
is that it can lead to undesired suboptimal solutions [47], [48].
Algorithm. 1 Channel estimation for STBC systems using
classical SD algorithm
1: compute R
2: perform the eigenvalue decomposition R = UU
H
3: compute the whitened data X

with (9)
4: initialize W randomly
5: repeat
6: set

(W)
7: compute the gradient
W
in the Euclidean space with
(26)
8: update W W
W
9: update W W(W
H
W)
1/2
10: until

(W) < where is a threshold

11: compute

H with (8).
Recently, major improvements have been obtained by tak-
ing into account the geometrical aspect of the optimization
problem. Nongeodesic and geodesic approaches have been
proposed in [48], [49]. Coupled with Armijo step size rule
[50], these algorithms always converge to a local minimum
if it is not initialized at a stationary point. In the following
equations, the geodesic SD algorithm [48] is chosen since it
has lower computational complexity than the nongeodesic one.
The geodesic SD algorithm moves towards the SD gradient
direction,
W
, in the Riemannian space. This direction can
be expressed as [48]

W
=
W
W
H
W
H
W
(29)
where
W
is the gradient in the Euclidean space (see (26)).
Then, the update rule is given by
W exp (
W
) W (30)
where exp() =

=0
()

/! is the matrix exponential and

corresponds to the step size. Using the Armijo step size rule,
the algorithm almost always converges to a local minimum.
The geodesic SD algorithm with the Armijo step size rule is
described in the algorithm 2.
Figures 1 and 2 illustrate the convergence of algorithm 2 for
a STBC system. The STBC system employs Alamouti coding
2
As discussed in [47], line search optimization is not well-adapted for
Euclidean SD with the projection method.
h
t
t
p
:
/
/
w
w
w
.
i
e
e
e
x
p
l
o
r
e
p
r
o
j
e
c
t
s
.
b
l
o
g
s
p
o
t
.
c
o
m
h
t
t
p
:
/
/
w
w
w
.
i
e
e
e
x
p
l
o
r
e
p
r
o
j
e
c
t
s
.
b
l
o
g
s
p
o
t
.
c
o
m
CHOQUEUSE et al.: BLIND CHANNEL ESTIMATION FOR STBC SYSTEMS USING HIGHER-ORDER STATISTICS 499
Algorithm. 2 Channel estimation for STBC systems using
geodesic SD algorithm with Armijo step size rule
1: compute R
2: perform the eigenvalue decomposition R = UU
H
3: compute the whitened data X

with (9)
4: initialize W randomly and set = 1
5: repeat
6: compute the gradient
W
in the Euclidean space with
(26)
7: compute the direction
W
in the Riemannian space
with (29)
8: compute
W
,
W
=
1
2

trace(
W

H
W
)

9: initialize = exp(
W
) and = .
10: while (W) (W)
W
,
W
do
11: set = , = and 2
12: end while
13: while (W) (W) <

2

W
,
W
do
14: set = exp(
W
) and
1
2

15: end while

16: update W W
17: until
W
,
W
< where is a threshold
18: compute

H with (8).
0 1 2 3 4 5 6 7 8 9
2.2
2
1.8
1.6
1.4
1.2
1
0.8
0.6
0.4
Number of iterations
C
o
s
t

f
u
n
c
t
i
o
n


K[ s1(t)]
K[ s2(t)]
Cost Function: J(W)
Fig. 1: Cost function (W) versus iteration.
(

= = = 2) and QPSK modulation. The number of

transmitted blocks, the number of receiver antennae and the
Signal to Noise Ratio (SNR) are equal to

= 512,

= 4
and 20dB, respectively, and the threshold is xed at = 10
5
.
Figure 1 displays [
1
], [
2
] and (W) with respect to the
iteration number. The gure shows that the cost function is
minimized after 9 iterations. The kurtosis [
1
] and [
2
]
converge to 1 which is the kurtosis of QPSK modulation
[38]. Figure 2 shows the constellation of the symbols
1
and
2
in the complex plane before and after convergence.
After convergence, it should be noted that the constellation
of the equalized symbols is phase-rotated as compared to the
QPSK constellation. However, as opposed to the classical ICA
model, the phase rotation ambiguities of
1
and
2
are not
independent. The effect of the STBC structure on the channel
ambiguities is studied in the following section.
1.5 1 0.5 0 0.5 1 1.5
1.5
1
0.5
0
0.5
1
1.5
Q
u
a
d
r
a
tu
r
e
InPhase
(a) After Whitening:
1
1.5 1 0.5 0 0.5 1 1.5
1.5
1
0.5
0
0.5
1
1.5
Q
u
a
d
r
a
tu
r
e
InPhase
(b) After Whitening:
2
1.5 1 0.5 0 0.5 1 1.5
1.5
1
0.5
0
0.5
1
1.5
Q
u
a
d
r
a
tu
r
e
InPhase
(c) Iteration 9:
1
1.5 1 0.5 0 0.5 1 1.5
1.5
1
0.5
0
0.5
1
1.5
Q
u
a
d
r
a
tu
r
e
InPhase
(d) Iteration 9:
2
Fig. 2: Constellation of
1
and
2
V. REMAINING AMBIGUITIES
For the classical ICA model, it is well known that the
channel can be estimated up to a permutation and phase
rotation ambiguities [23], [51]. For STBC systems, the pro-
posed method reduces the number of channel ambiguities by
exploiting the spatio-temporal redundancy of the transmitted
symbols in (18).
Theorem 1: Let us consider three matrices M, P and D
where M is an

unitary matrix, P is an permutation

matrix and D is an diagonal matrix with entries of unit
modulus (DD

= I

). If these matrices satisfy

M(s

) = (PDs

) (31)
for any s

, then HM
H
is also a solution of the blind channel
estimation problem.
Proof: From (4) and (31), one gets:
Y

= H(s

) + B

= HM
H
M(s

) + B

= HM
H
(PDs

) + B

. (32)
As the elements of s

are i.i.d, the elements of the vector

PDs

are also i.i.d. Therefore HM

H
is a solution of the blind
channel estimation problem.
Let us express the set , which contains all the matrices M
satisfying (31), with respect to the coding matrices. Condition
(31) can be described in a vector form as

(vec{
H
(s

)M
H
})
(vec{
H
(s

)M
H
})

. .. .
k
1
=

(vec{
H
(PDs

)})
(vec{
H
(PDs

)})

. .. .
k
2
.
(33)
The 2

-dimensional column vector vec{

H
(s

)M
H
} can be
expressed as
vec{
H
(s

)M
H
} = (M

) vec{
H
(s

)} (34)
h
t
t
p
:
/
/
w
w
w
.
i
e
e
e
x
p
l
o
r
e
p
r
o
j
e
c
t
s
.
b
l
o
g
s
p
o
t
.
c
o
m
h
t
t
p
:
/
/
w
w
w
.
i
e
e
e
x
p
l
o
r
e
p
r
o
j
e
c
t
s
.
b
l
o
g
s
p
o
t
.
c
o
m
500 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 10, NO. 2, FEBRUARY 2011
TABLE I: Set of ambiguity matrices for different STBCs using

= 2, 3, 4 transmit antennae. The matrices M

1
-M
10
are
dened in (40)-(49).
Number of Number of
Design transmit symbols Code rate Set of ambiguity matrices
approach antennae per block / after channel estimation

Spatial Multiplex.

= {PD}
Alamouti [1] 2 2 1 = {M
1
(), M
2
()}
OSTBC [2] 3 3 1/2 = {I
3
}
OSTBC [32] 3 3 3/4 = {I
3
}
OSTBC [3] 4 3 3/4 = {I
4
}
OSTBC [2] 4 4 1/2 = {I
4
, M
4
(0), M
5
(0), M
6
(0)}
QOSTBC [4] 4 4 1 = {M
3
(), M
4
(), M
5
(), M
6
()}
NOSTBC [52] 4 4 1 = {M
7
(
1
,
2
), M
8
(
1
,
2
), M
9
(
1
,
2
), M
10
(
1
,
2
)}
As (M

) = (M) and (M

) = (M), k
1
can be
written in a linear form as
k
1
=

(M) I

(M) I

(M) I

(M) I

Gs

. (35)
The right term in (33) can also be expressed into a linear form
as
k
2
= G

(vec{PDs

})
(vec{PDs

})

= G

P(D) P(D)
P(D) P(D)

. (36)
Using (35) and (36), (33) can be simplied as
M
T
Gs

= G

P(D) P(D)
P(D) P(D)

(37)
where M
T
is an 2

matrix with real elements, which

is dened as
M
T
=

(M) I

(M) I

(M) I

(M) I

. (38)
As (37) must be satised for any s

, one obtains
G

M
T
G =

P(D) P(D)
P(D) P(D)

. (39)
Finally, the following result is obtained
Theorem 2: For any STBC , the set of ambiguity
matrices is the one containing all the

matrices M
satisfying (39) where P is a

permutation matrix
and D is a

diagonal matrix with entries of unit modulus.

It should be noted that the condition (39) depends on
the matrix G which only depends on the STBC (see (15)).
Unfortunately, it appears to be difcult to nd the exact
relationship between G and the matrices M, P and D. To
provide a clear relationship between these matrices, we have
performed several Monte-Carlo simulations with the Rayleigh
MIMO channel. Table I provides the set of ambiguity
matrices for several STBCs using

= {2, 3, 4} transmit
antennae. In Table I, matrices M
1
-M
10
are equal to
M
1
() =

0
0

(40)
M
2
() =

(41)
M
3
() =

M
1
() 0
2
0
2
M
1
()

(42)
M
4
() =

M
2
() 0
2
0
2
M
2
()

(43)
M
5
() =

0
2
M
1
()
M
1
() 0
2

(44)
M
6
() =

0
2
M
2
()
M
2
() 0
2

(45)
M
7
(
1
,
2
) =

M
1
(
1
) 0
2
0
2
M
1
(
2
)

(46)
M
8
(
1
,
2
) =

M
1
(
1
) 0
2
0
2
M
2
(
2
)

(47)
M
9
(
1
,
2
) =

M
2
(
1
) 0
2
0
2
M
1
(
2
)

(48)
M
10
(
1
,
2
) =

M
2
(
1
) 0
2
0
2
M
2
(
2
)

. (49)
Let us emphasize the differences between Table I and the
tables reported in [17], [18], [21], [22]. Tables reported in [17],
[18], [21], [22] focus on the blind channel-identiability con-
dition for subspace and SOS approaches. Without modication
of the transmitter (precoding, pilot sequence), they show that
subspace and SOS methods are unable to estimate the channel
for 1-rate STBCs and some specic low-rate STBCs. Unlike
subspace and SOS approaches, the proposed method can be
applied to the whole class of linear STBCs without any
modication of the transmitter or the use of a pilot sequence.
Moreover, unlike the general subspace methods [16], [17], it
does not introduce additional ambiguities to those associated
to the blind channel estimation problem. For example for the
3
4
-rate OSTBC using

= 3 antennae, the proposed method

can estimate the channel up to a sign whereas the subspace
method introduces an unknown phase rotation

[16], [17].
VI. SIMULATION RESULTS
Monte-Carlo simulations were run to assess the perfor-
mances of the algorithms 1 and 2. Let us denote by H
and

H the original and estimated channel, respectively. After
h
t
t
p
:
/
/
w
w
w
.
i
e
e
e
x
p
l
o
r
e
p
r
o
j
e
c
t
s
.
b
l
o
g
s
p
o
t
.
c
o
m
h
t
t
p
:
/
/
w
w
w
.
i
e
e
e
x
p
l
o
r
e
p
r
o
j
e
c
t
s
.
b
l
o
g
s
p
o
t
.
c
o
m
CHOQUEUSE et al.: BLIND CHANNEL ESTIMATION FOR STBC SYSTEMS USING HIGHER-ORDER STATISTICS 501
TABLE II: Spatial Multiplexing: Average Computation times
for each algorithm.

Algorithm
SNR
-10dB 0dB 10dB
Classical SD 0.07 s 0.11 s 0.09 s
Geodesic SD 0.34 s 0.16 s 0.09 s
JADE 0.004 s 0.004 s 0.004 s
channel estimation, the remaining ambiguity is removed by
post-multiplying

H with

M where

M = arg min
M
H

HM
2

(50)
and where the set of ambiguity matrices, , depends on the
STBC (see Table I). After ambiguity removal, the estimated
channel is denoted as

H

=

H

M. Performances of the
proposed blind algorithms were quantied through:
the Normalized Mean Square Error (NMSE), which is
dened as:
=
H

H

H
2

. (51)
the average Symbol Error Rate (SER) obtained after ML
decoding.
Each simulation was carried out under the following condi-
tions: i) a Rayleigh distributed channel i.e. each element of
H follows an i.i.d. circular Gaussian distribution with zero-
mean and unit-variance, ii) a QPSK modulation, iii) 512
transmitted blocks, iv) a temporally and spatially zero-mean
white Gaussian additive noise with variance
2
(which is
unknown at the receiver side), v) a threshold equal to = 10
5
and vi) a receiver satisfying assumption AS1). Performances
of the algorithms 1 and 2 were evaluated for several Signal-
to-Noise Ratios (SNRs) where the SNR was dened as [53]
= 10log
10
(

/
2
). (52)
For each SNR, two thousand Monte-Carlo simulations were
performed to approximate the NMSE and SER. As there is
no guarantee that the algorithms 1 and 2 will nd the global
minimum, performances of the proposed methods were also
evaluated with multistart initialization. Multistart initialization
procedure runs an algorithm several times with new random
starting points and selects the estimated unitary matrix W
which minimizes the cost-function (W). In the following
subsections, performances are presented for 3 different STBC
systems.
A. Spatial Multiplexing
In this subsection, we consider the case of a Spatial
Multiplexing system using

= 2 transmit antennae. The

transmitted blocks are given by
(s) =

. (53)
For Spatial Multiplexing, the channel estimation problem re-
duces to the classical ICA problem. After channel estimation,
the set of ambiguity matrices is given by = {PD} where
10 5 0 5 10
10
3
10
2
10
1
10
0
SNR (dB)
N
M
S
E


Proposed Method: Classical SD
Proposed Method: Classical SD (5 starts)
Proposed Method: Geodesic SD
Proposed Method: Geodesic SD (5 starts)
JADE
Fig. 3: Spatial Multiplexing: NMSE.
10 5 0 5 10
10
2
10
1
10
0
SNR (dB)
S
E
R


Coherent receiver
Proposed Method: Classical SD
Proposed Method: Classical SD (5 starts)
Proposed Method: Geodesic SD
Proposed Method: Geodesic SD (5 starts)
JADE
Fig. 4: Spatial Multiplexing: Symbol Error Rate.
P and D are permutation and phase matrices. Figures 3 and
4 present the performances of the algorithms 1 and 2 for a
receiver composed of

= 3 antennae. These two algorithms

are compared with JADE [54]. Figure 3 displays the channel
NMSE versus SNR. In this simulation, algorithms 1 and 2
always match or outperform the JADE algorithm, depending
on the SNR. Figure 3 also indicates that the multistart initial-
ization does not seem to improve the performances. Figure 4
presents the SER versus the SNR. The SER is compared to the
one obtained with the coherent ML receiver (perfect CSI). It
should be noted that the blind channel-estimation algorithms
achieve near-optimal performances at high SNR since their
SERs approach the ones of the coherent ML receiver. A
comparison of the average computation times is shown in
Table II for simulations implemented on a 2.6 GHz Intel
Pentium processor using Matlab. For multistart initialization
approaches, the computation times must be multiplied by the
number of random starts. Table II shows that the classical
SD is less computationally demanding than the geodesic SD
at low-SNR, but their computational complexities are similar
at high-SNR. Table II also suggests that the JADE algorithm
is signicantly less complex than the proposed algorithms.
h
t
t
p
:
/
/
w
w
w
.
i
e
e
e
x
p
l
o
r
e
p
r
o
j
e
c
t
s
.
b
l
o
g
s
p
o
t
.
c
o
m
h
t
t
p
:
/
/
w
w
w
.
i
e
e
e
x
p
l
o
r
e
p
r
o
j
e
c
t
s
.
b
l
o
g
s
p
o
t
.
c
o
m
502 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 10, NO. 2, FEBRUARY 2011
10 5 0 5 10
10
4
10
3
10
2
10
1
10
0
10
1
SNR (dB)
N
M
S
E


Proposed Method: Classical SD
Proposed Method: Classical SD (5 starts)
Proposed Method: Geodesic SD
Proposed Method: Geodesic SD (5 starts)
Fig. 5: Alamouti Coding: NMSE.
10 5 0 5 10
10
4
10
3
10
2
10
1
10
0
SNR (dB)
S
E
R


Coherent receiver
Proposed Method: Classical SD
Proposed Method: Classical SD (5 starts)
Proposed Method: Geodesic SD
Proposed Method: Geodesic SD (5 starts)
Fig. 6: Alamouti Coding: Symbol Error Rate.
TABLE III: Alamouti Coding: Average Computation times for
each algorithm.

Algorithm
SNR
-10dB 0dB 10dB
Classical SD 0.10 s 0.16 s 0.12 s
Geodesic SD 0.51 s 0.21 s 0.10 s
However, one should note that JADE is limited to Spatial-
Multiplexing systems and cannot be employed for more gen-
eral settings.
B. Alamouti Coding
In this subsection, we consider a STBC system using the
Alamouti Code. This Orthogonal code is dened by [1]
(s) =

(54)
For Alamouti coding, the direct use of an ICA algorithm,
like JADE, is irrelevant since the transmitted symbols between
consecutive time instances are not independent. Furthermore,
it is demonstrated in [16][18], [22] that subspace and SOS
approaches cannot estimate the channel matrix when the
transmitter employs Alamouti Coding. Regarding the proposed
methods, Table I shows that the set of ambiguity matrices after
channel estimation is = {M
1
(), M
2
()}. Figure 5 displays
the NMSE versus SNR for a receiver composed of

= 3
antennae. Without multistart initialization, the geodesic SD
clearly outperforms the classical SD since the latter exhibits
an error oor at SNR greater than 4dB. This error oor is
due to the fact that the Euclidean SD can lead to undesired
suboptimal solutions even at high SNR [47], [48]. It should
be observed that the multistart initialization strategy removes
the error oor and improves the NMSE performances of the
two proposed algorithms. Figure 6 compares the SER with
the one obtained with a coherent ML receiver. As previously
discussed, without multistart initialization, the performances
of the Euclidean SD lead to an error oor at SNR greater
than 4dB. However, it should be observed that algorithms 1
and 2 achieve near-optimal performance when a multistart ini-
tizalization is used. A comparison of the average computation
times is shown in Table III. It should be noted that classical
SD is less computationally demanding than the geodesic SD
at low-SNR, but this trend is reversed at high SNR.
C.
3
4
-rate OSTBC using 3 antennae
In this subsection, we consider the case of a
3
4
-rate OSTBC
using 3 antennae. This OSTBC is dened by [32]
(s) =

1
0
2

3
0
1

2

3

1
0

(55)
For this low-rate code, the channel can be estimated with
subspace and SOS approaches. The remaining ambiguity
reduces to a sign for the SOS approach [18] and to a rotation

for the subspace method [16], [17]. In the following g-

ures, performances of the proposed algorithms are compared
with the SOS-based method [18] for a receiver composed of

= 5 antennae. Figure 7 displays the NMSE versus SNR.

Without multiple-start initialization, it should be noted that
the algorithms 1 and 2 exhibit an error oor at SNR greater
than 6dB. For the Euclidean SD algorithm, the error oor
is due to the fact that this approach can lead to undesired
suboptimal solutions [47], [48]. For the Geodesic SD one,
even if it always converges to a local minimum [47], the
error oor is due to the fact that the local minimum does
not necessarily coincide with the global one. As previously
observed, multistart initialization removes the error oor and
improves the performances of the two proposed algorithms. In
particular, Figure 7 shows that for > 0, algorithms 1 and
2 with multistart initialization outperform the SOS method.
Figure 8 shows that method [18] and the proposed multistart
algorithms both achieve near-optimal SER performances for
> 0. A comparison of the average computation times
is presented in Table IV. It is shown that the SOS method is
less computationally demanding than the proposed algorithms.
Therefore, for SOS-identiable OSTBCs, it seems that the
closed-form SOS algorithm is denitely preferable since this
algorithm does not suffer from convergence problems and it is
less computationally expensive than the proposed approaches.
However, it should be emphasized that the SOS method is
h
t
t
p
:
/
/
w
w
w
.
i
e
e
e
x
p
l
o
r
e
p
r
o
j
e
c
t
s
.
b
l
o
g
s
p
o
t
.
c
o
m
h
t
t
p
:
/
/
w
w
w
.
i
e
e
e
x
p
l
o
r
e
p
r
o
j
e
c
t
s
.
b
l
o
g
s
p
o
t
.
c
o
m
CHOQUEUSE et al.: BLIND CHANNEL ESTIMATION FOR STBC SYSTEMS USING HIGHER-ORDER STATISTICS 503
10 5 0 5 10
10
4
10
3
10
2
10
1
10
0
SNR (dB)
N
M
S
E


Proposed Method: Classical SD
Proposed Method: Classical SD (5 starts)
Proposed Method: Geodesic SD
Proposed Method: Geodesic SD (5 starts)
SOS Method
Fig. 7: 3/4-rate OSTBC: NMSE.
10 5 0 5 10
10
6
10
5
10
4
10
3
10
2
10
1
10
0
SNR (dB)
S
E
R


Coherent receiver
Proposed Method: Classical SD
Proposed Method: Classical SD (5 starts)
Proposed Method: Geodesic SD
Proposed Method: Geodesic SD (5 starts)
SOS Method
Fig. 8: 3/4-rate OSTBC: Symbol Error Rate.
TABLE IV:
3
4
-rate OSTBC: Average Computation times for
each algorithm.

Algorithm
SNR
-10dB 0dB 10dB
Classical SD 1.32 s 0.85 s 0.78 s
Geodesic SD 3.33 s 0.62 s 0.51 s
SOS method 0.48 s 0.48 s 0.48 s
limited to a subclass of OSTBCs [22], whereas the proposed
algorithms can be applied to the whole class of linear STBCs.
VII. CONCLUSION
This paper proposed an original blind channel estimation
algorithm for space-time block coding communications. The
method is based on the minimization of a kurtosis-based
cost function after Zero-Forcing equalization. The proposed
method can be applied to the whole class of linear STBCs,
whatever the code-rate and the modulation. This paper also
presented the set of the remaining channel ambiguities for
several STBCs using 2, 3 or 4 transmit antennae. The good
performances of the proposed algorithm were demonstrated
through computer simulations for different STBCs. In par-
ticular, simulations have shown that the proposed method
matches or outperforms the JADE algorithm [54] for Spatial
Multiplexing and matches the performances of the closed-form
SOS approach [18] for identiable OSTBCs.
APPENDIX
Let us consider the 2

matrix W dened in (12).

As (A B)
T
= A
T
B
T
, one gets:
W
T
W =

(W) I

(W) I

(W) I

(W) I

(W
T
) I

(W
T
) I

(W
T
) I

(W
T
) I

(56)
From the mixed product rule, it follows that:
W
T
W =

B
1
I

B
2
I

B
2
I

B
1
I

(57)
where the

matrices B
1
and B
2
are given by:
B
1
= (W)(W
T
) +(W)(W
T
) (58)
B
2
= (W)(W
T
) (W)(W
T
) (59)
As W is a

unitary matrix, it satises WW

H
= I

.
By expanding the real and imaginary parts, one gets:
WW
H
=

(W) +(W)

(W
T
) (W
T
)

= B
1
B
2
= I

(60)
By identication, it follows that B
1
= I

and B
2
= 0

.
Finally (57) can be simplied as:
W
T
W = I
2

(61)
REFERENCES
[1] S. Alamouti, A simple transmit diversity technique for wireless com-
munication," IEEE J. Sel. Areas Commun., vol. 16, no. 8, pp. 1451-1458,
1998.
[2] V. Tarokh, H. Jafarkhani, and A. Calderbank, Space time block codes
from orthogonal designs," IEEE Trans. Inf. Theory, vol. 45, no. 5, pp.
744-765, 1999.
[3] G. Ganesan and P. Stoica, Space-time block codes: a maximum SNR
approach," IEEE Trans. Inf. Theory, vol. 47, no. 4, pp. 1650-1656, 2001.
[4] H. Jafarkhani, A quasi-orthogonal space-time block code," IEEE Trans.
Commun., vol. 49, no. 1, pp. 1-4, 2001.
[5] A. Boarui and D. Ionescu, A class of nonorthogonal rate-one space-
time block codes with controlled interference," IEEE Trans. Wireless
Commun., vol. 2, no. 2, pp. 270-276, 2003.
[6] H. Jafarkhani, Space-Time Coding: Theory and Practice. Cambridge
University Press, 2005.
[7] G. Ganesan and P. Stoica, Differential modulation using space-time
block codes," IEEE Signal Process. Lett., vol. 9, no. 2, pp. 57-60, 2002.
[8] B. Hochwald and W. Sweldens, Differential unitary space-time modu-
lation," IEEE Trans. Commun., vol. 48, no. 12, pp. 2041-2052, 2000.
[9] B. Hughes, Differential space-time modulation," IEEE Trans. Inf.
Theory, vol. 46, no. 7, pp. 2567-2578, 2000.
[10] Y. Zhu and H. Jafarkhani, Differential modulation based on quasi-
orthogonal codes," IEEE Trans. Wireless Commun., vol. 4, no. 6, pp.
3005-3017, 2005.
[11] E. Larsson, P. Stoica, and J. Li, On maximum-likelihood detection and
decoding for space-time coding systems," IEEE Trans. Signal Process.,
vol. 50, no. 4, pp. 937-944, 2002.
[12] , Orthogonal space-time block codes: maximum likelihood de-
tection for unknown channels and unstructured intereferences," IEEE
Trans. Signal Process., vol. 51, no. 2, pp. 362-372, 2003.
[13] W. Ma, B. Vo, T. Davidson, and P. Ching, Blind ML detection of
orthogonal space-time block codes: efcient high-performance imple-
mentations," IEEE Trans. Signal Process., vol. 54, no. 2, pp. 738-751,
2006.
h
t
t
p
:
/
/
w
w
w
.
i
e
e
e
x
p
l
o
r
e
p
r
o
j
e
c
t
s
.
b
l
o
g
s
p
o
t
.
c
o
m
h
t
t
p
:
/
/
w
w
w
.
i
e
e
e
x
p
l
o
r
e
p
r
o
j
e
c
t
s
.
b
l
o
g
s
p
o
t
.
c
o
m
504 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 10, NO. 2, FEBRUARY 2011
[14] Y. Li, C. Georghiades, and G. Huang, Iterative maximum likelihood
sequence estimation for space-time coded systems," IEEE Trans. Com-
mun., vol. 49, no. 6, pp. 948-951, 2001.
[15] A. Gallo, E. Chiavaccini, F. Muratori, and G. Vitetta, BEM-based SISO
detection of orthogonal space-time block codes over frequency at-
fading channels," IEEE Trans. Wireless Commun., vol. 3, no. 6, pp.
1885-1889, 2004.
[16] A. Swindlehurst and G. Leus, Blind and semi-blind equalization for
generalized space-time block codes," IEEE Trans. Signal Process.,
vol. 50, no. 10, pp. 2489-2498, 2002.
[17] N. Ammar and Z. Ding, Blind channel identiability for generic linear
space-time block codes," IEEE Trans. Signal Process., vol. 55, no. 1,
pp. 202-217, 2007.
[18] S. Shahbazpanahi, A. Gershman, and J. Manton, Closed form blind
MIMO channel estimation for othogonal space-time codes," IEEE Trans.
Signal Process., vol. 53, no. 12, pp. 4506-4517, 2005.
[19] J. Via, I. Santamaria, A. Sezdin, and A. Paulraj, SOS-based blind chan-
nel estimation algorithm under space-time block coded transmissions,"
in Eighth IEEE Workshop Signal Process. Advances Wireless Commun.,
Helsinki, Finland, June 2007.
[20] J. Via and I. Santamaria, Correlation matching approaches for blind
OSTBC channel estimation," IEEE Trans. Signal Process., vol. 56,
no. 12, pp. 5950-5961, 2008.
[21] J. Via, I. Santamaria, and J. Perez, Code combination for blind channel
estimation in general MIMO-STBC systems," EURASIP J. Advances
Signal Process., 2009.
[22] J. Via and I. Santamaria, On the blind identiablity of orthogonal space
time block codes from second order statistics," IEEE Trans. Inf. Theory,
vol. 54, no. 2, pp. 709-722, 2008.
[23] A. Hyvarinen, J. Karunen, and E. Oja, Independent Component Analysis.
John Wiley and Sons, 2001.
[24] B. Gu, J. Liu, and Y. Yu, Orthogonal detection of beam space time
block coding using ICA," in IEEE Neural Netw. Brain, vol. 2, Beijing,
China, Oct. 2005, pp. 836-840.
[25] J. Via, I. Santamaria, and J. Peres, Blind identication of MIMO-
OSTBC channels combining second and higher order statistics," in
European Signal Process. Conf., EUSIPCO, Florence, Italy, Sep. 2006.
[26] E. Beres and R. Adve, Blind channel estimation for orthogonal STBC
in MISO systems," IEEE Trans. Veh. Technol., vol. 56, no. 4, pp. 2042-
2050, 2007.
[27] S. Daumont and D. L. Guennec, Blind source separation with order
recovery for MIMO system and an Alamouti or Tarokh space-time block
coding scheme," in Proc. IEEE International Symp. Signal Process. Inf.
Technol., Cairo, Egypt, Dec. 2007, pp. 431-436.
[28] H. Iglesias, J. Garcia-Naya, and A. Dapena, A blind channel estimation
strategy for the 2x1 Alamouti system based on diagonalising 4th order
cumulant matrices," in Proc. International Conf. Acoustic Speech Signal
Process., Las Vegas, USA, Mar. 2008, pp. 3329-3332.
[29] A. Mansour, J. Youssef, and K. Yao, Underdetermined BSS of MISO
OSTBC signals," in ICA, ser. Lecture Notes Comput. Science, vol. 5441,
Paraty, Brazil, Mar. 2009, pp. 678-685.
[30] H. Xu, J. Liu, A. Perez-Neira, and M. Lagunas, Independent component
analysis applied to multiple antenna space-time systems," in Proc. IEEE
16th International Symp. Personal, Indoor Mobile Radio Commun.,
Berlin, Germany, 2005, pp. 57-61.
[31] J. Liu, A. Iserte, and M. Lagunas, Blind separation of OSTBC
signals using ICA neural networks," in Proc. IEEE ISSPIT, Darmstadt,
Germany, Dec. 2003, pp. 502-505.
[32] E. Larsson and P. Stoica, Space-Time Block Coding for Wireless Com-
munication. Cambridge University Press, 2003.
[33] M. Shi, Y. Bar-Ness, and W. Su, STC and BLAST MIMO modulation
recognition," in Proc. IEEE GLOBECOM, Nov. 2007, pp. 3034-3039.
[34] V. Choqueuse, K. Yao, L. Collin, and G. Burel, Blind recognition of
linear space time block codes," in Proc. IEEE ICASSP, Las Vegas, USA,
Mar. 2008, pp. 2833-2836.
[35] , Hierarchical space time block code recognition using correlation
matrices," IEEE Trans. Wireless Commun., vol. 7, no. 9, pp. 3526-3534,
2008.
[36] V. Choqueuse, M. Marazin, L. Collin, K. Yao, and G. Burel, Blind
recognition of linear space-time block codes: a likelihood-based ap-
proach," IEEE Trans. Signal Process., vol. 58, no. 3, pp. 1290-1299,
2010.
[37] J. Brewer, Kronecker products and matrix calculus in system theory,"
IEEE Trans. Circuits Syst., vol. 25, no. 9, pp. 772-781, 1978.
[38] A. Swami and B. Sadler, Hierarchical digital modulation classication
using cumulants," IEEE Trans. Commun., vol. 48, no. 3, pp. 416-429,
2000.
[39] N. Delfosse and P. Loubaton, Adaptive separation of independent
sources: a deation approach," in Proc. IEEE International Conf.
Acoustic, Speech Signal Process., Adelaide, Australia, 1994, pp. 41-44.
[40] J. Cardoso and B. Laheld, Equivariant adaptive source separation,"
IEEE Trans. Signal Process., vol. 44, no. 12, pp. 3017-3030, 1996.
[41] E. Moreau and O. Macchi, High order contrasts for self-adaptive source
separation," International J. Adaptive Control Signal Process., vol. 10,
no. 1, pp. 19-46, 1996.
[42] E. Moreau, Criteria for complex sources separation," in Proc. EU-
SIPCO, Trieste, Italy, 1996, pp. 931-934.
[43] H. Li and T. Adali, A class of complex ICA algorithms based on the
kurtosis cost function," IEEE Trans. Neural Netw., vol. 19, no. 3, pp.
408-420, 2008.
[44] D. Brandwood, A complex gradient operator and its application in
adaptive array theory," IEE Process., Parts F and H, vol. 130, no. 1,
pp. 11-16, 1983.
[45] LEcuyer, Note: on the interchange of derivative and expectation for
likelihood ratio derivative estimators," Management Science, vol. 41,
no. 4, pp. 738-747, 1995.
[46] C. Papadias, Globally convergent blind source separation based on a
multiuser kurtosis maximisation criterion," IEEE Trans. Signal Process.,
vol. 48, no. 12, pp. 3508-3519, 2000.
[47] T. Abrudan, Advanced optimization algorithms for sensor arrays and
multi-antenna communications," Ph.D. dissertation, Helsinky University
of Technology, 2008.
[48] T. Abrudan, J. Eriksson, and V. Koivunen, Steepest descent algorithms
for optimization under unitary matrix constraint," IEEE Trans. Signal
Process., vol. 56, no. 3, pp. 1134-1147, 2008.
[49] J. Manton, Optimization algorithms exploiting unitary constraints,"
IEEE Trans. Signal Process., vol. 50, no. 3, pp. 635-650, 2004.
[50] E. Polak, Optimization: Algorithms and Consistent Approximations.
Springer-Verlag, 1997.
[51] A. Mansour and M. Kawamoto, ICA papers classied according to their
applications and performances," IEICE Trans. Fundamentals Electron.,
Commun. Comput. Sciences, vol. E86, no. A, pp. 620-633, 2000.
[52] IEEE 802.16e, standard for local and metropolitan area networks, part
16: air interface for xed and mobile broadband wireless access system,"
2005.
[53] B. Vucetic and J. Yuan, Space-Time Coding. John Wiley & Sons, 2003.
[54] J. Cardoso and A. Soloumiac, Blind beamforming for non-Gaussian
signals," in IEE Proc. F, vol. 140, no. 46, pp. 362-370, 1993.
Vincent Choqueuse (S08-M09) was born in 1981
in Brest, France. He received the Dipl.-Ing. and the
M.Sc. degrees in 2004 and 2005, respectively, from
Troyes University of Technology (UTT), France,
and the Ph.D. degree in 2008 from University of
Brest, France. Since September 2009, he has been
Associate Professor at the IUT of Brest, France,
and a member of the Laboratory LBMS (EA 4325).
His research interests focus on signal processing and
statistics for communications and diagnosis.
Ali Mansour (M97,SM00) was born at Tripoli
in Lebanon in 1969. He received his Electronic-
Electrical Engineering Diploma in 1992 from the
Lebanese University, Tripoli, Lebanon, and his
M.Sc. and the Ph.D. degrees in Signal, Image and
Speech Processing from INPG, Grenoble, France, in
1993 and 1997, respectively. From January 1997 to
July 1997, he held a post doc position at LTIRF-
INPG, Grenoble, France. From August 1997 to
September 2001, he was a Research Scientist at
the Bio-Mimetic Control Research Center of Riken,
Nagoya, Japan. From 2001 to 2008, he was holding a teacher-researcher
position at the Ecole Nationale Suprieure des Ingnieurs des Etudes et
Techniques dArmement (ENSIETA),Brest, France. Since February 2008, he
has been a senior-lecturer at the Department of Electrical and Computer
Engineering at Curtin University of Technology, Perth, Australia. His research
interests are in the areas of blind separation of sources, high order statistics,
signal processing, COMINT, radar, sonar and robotics.
h
t
t
p
:
/
/
w
w
w
.
i
e
e
e
x
p
l
o
r
e
p
r
o
j
e
c
t
s
.
b
l
o
g
s
p
o
t
.
c
o
m
h
t
t
p
:
/
/
w
w
w
.
i
e
e
e
x
p
l
o
r
e
p
r
o
j
e
c
t
s
.
b
l
o
g
s
p
o
t
.
c
o
m
CHOQUEUSE et al.: BLIND CHANNEL ESTIMATION FOR STBC SYSTEMS USING HIGHER-ORDER STATISTICS 505
Ludovic Collin received the Ph.D. degree in elec-
trical engineering from the University of Bretagne
Occidentale, Brest, France, in 2002. From 1989 to
1999 he was with ORCA Instrumentation, Brest,
where he developed oceanographical instrumenta-
tion and acoustic modems. From 1999 to 2002
he was Research and Teaching Assistant at French
Naval Academy, Lanveoc, France, and at the In-
stitute of Technology of Lannion. From 2003 to
2007 he was Assistant Professor at the ENSIETA,
Brest. Since 2007 he has been Assistant Professor
at the University of Brest and member of the Laboratory for Science and
Technologies of Information, Communication and Knowledge (Lab-STICC
- UMR CNRS 3192). His research interests are in MIMO systems and
interception of communications.
Kof Clment Yao (M05) received the PhD degree
in Optical Signal Processing and computer sciences
from University Louis Pasteur of Strasbourg, France
in 1990. After his post-doctorate research on optical
neural networks at Ecole Nationale Suprieure des
Tlcommunications of Brest, he joined the French
naval academy as assistant professor in statistical
signal processing in 1992. His research interest was
focused on Pattern recognition and blind signal
separation in underwater acoustics. Since 2001, he
has been Assistant Professor at University of Brest,
France. His present research interests are in MIMO systems and blind
interception of digital communication signals.
Gilles Burel (M00-SM08) was born in 1964. He
received the M.Sc. degree from Ecole Suprieure
dElectricit, Gif Sur Yvette, France, in 1988, the
Ph.D. degree from University of Brest, France, 1991,
and the Habilitation to Supervise Research degree
in 1996. From 1988 to 1997 he was a member of
the technical staff of Thomson CSF, then Thomson
Multimedia, Rennes, France, where he worked on
image processing and pattern recognition applica-
tions as project manager.
Since 1997, he has been Professor of Digital
Communications, Image and Signal Processing at the University of Brest.
He is Associate Director of the Laboratory for Science and Technologies
of Information, Communication and Knowledge (Lab-STICC - UMR CNRS
3192). He is author or co-author of 19 patents, one book and 140 scientic
papers. His present research interests are in signal processing for digital
communications, MIMO systems and interception of communications.
h
t
t
p
:
/
/
w
w
w
.
i
e
e
e
x
p
l
o
r
e
p
r
o
j
e
c
t
s
.
b
l
o
g
s
p
o
t
.
c
o
m
h
t
t
p
:
/
/
w
w
w
.
i
e
e
e
x
p
l
o
r
e
p
r
o
j
e
c
t
s
.
b
l
o
g
s
p
o
t
.
c
o
m